AI developer tools · What shipped

Claude Fable 5 achieved a new high score of 161 on the Epoch Capabilities Index (ECI), surpassing GPT-5.5 Pro by 1 point and marking the first time Anthropic has led the benchmark in over a year. Epoch AI recently expanded its benchmarking hub by tracking 13 new evaluations as of July 1, 2026, with 7 incorporated into the ECI, alongside nine additional external benchmarks added on June 22 covering agentic work, cybersecurity, algorithm engineering, forecasting, and physics.

Moonshot AI released multiple AI developer tools and infrastructure updates including Kimi-Dev, a coding LLM achieving 60.4% on SWE-bench Verified, and Kimi Code, a fast AI coding agent for development workflows. Infrastructure releases include checkpoint-engine for efficiently updating model weights in LLM serving and FlashKDA, high-performance Kimi Delta Attention kernels. Research contributions shipped with code include Mooncake (KV-centric disaggregated LLM serving, Best Paper at FAST 2025), MoBA for long-context LLMs, and Moonlight with improved Muon optimizer for LLM training alongside SOTA small model checkpoints.

Unsloth released optimized quantizations and inference tools for Alibaba's new Qwen3.6 multimodal models (27B and 35B-A3B variants). The Qwen3.6-27B runs on 18GB RAM and 35B-A3B on 22GB, supporting 256K context across 201 languages with improvements in agentic coding and tool-calling. Unsloth's Dynamic 2.0 quantization approach calibrates on real-world datasets with selective layer upcasting, and they introduced Multi Token Prediction (MTP) speculative decoding enabling 1.4-2.2x faster generation on these models. On RTX 6000, Qwen3.6-27B MTP achieves 160 tokens/s and 35B-A3B reaches 240 tokens/s. New NVFP4 W4A4 quantizations for Blackwell GPUs (RTX 50X, B200, B300) run 2.5x faster than standard NVFP4 with comparable accuracy across MMLU-Pro, AIME 2025, and GPQA benchmarks. Models are available in Unsloth Studio (their open-source web UI), llama.cpp, and via vLLM/SGLang for deployment, with KL Divergence benchmarks showing Unsloth GGUFs achieving state-of-the-art performance on 21 of 22 quantization sizes.

Mistral released Robostral Navigate, an 8B embodied AI model for robotic navigation that achieves 76.6% success on R2R-CE unseen benchmarks using only a single RGB camera, outperforming multi-sensor approaches. The model was built entirely in-house using simulated data and token-efficient prefix-caching techniques that reduce training tokens by 22×, generalizes across wheeled, legged, and flying robots, and incorporates online reinforcement learning (CISPO algorithm) that improved performance by 3.2%.